Unhooding Claw - Oracle's latest Agentic Applications extension snaps into place

> We are entering an era where where we're moving away from AI doing simple work to really now making much more serious business impact for organizations. We're taking on more interesting, more complex work. So says Natalia Rachelson, SVP Application Development at Oracle as the firm today unleashes its latest agentic additions in the form of Fusion Claw. Before we get to the nitty gritty of what that entails, a quick reminder from Rachelson of the story so far: > Fusion applications have been on the AI journey for about two years now. We started with adding ChatGPT and other models came to market. We started with infusing Fusion with what we call generative AI use cases. Those were really cases of AI assisting with very specific tasks, mainly at that point in just writing content, so writing job descriptions, item descriptions, helping summarize performance reviews, very simple assistance. We have then introduced AI agents, a single task agent that can perform a pre-defined task. > > > We then introduced teams of agents where you had a supervisor agent and then a few specialist agents co-ordinating, working with each other, and the supervisor agent monitoring specialist agents' activities, just like a human manager would. From there, we introduced workflow agents. These were agents that could complete an entire or a sub-set of a workflow. > > > And at the end of March of this year, we announced Agentic Applications, a brand new class of applications, outcome-driven systems that were based again on teams of specialist agents working with each other, sharing context, and really getting work done in Fusion. While all of this has been happening, the nature of Fusion itself has evolved, Rachelson suggests: > Fusion originally was more of a system of record. Yes, it's automated a lot of things, but the steps were pretty much hardwired. We built in industry best practices and it was more of a monolith system that recorded what happened and assisted people as they did their work in something that you would call a system of record. > > > As we started to infuse Fusion with more and more AI, it has transformed itself into a system of outcomes. It's a paradigm shift, re-defining the contract between people and the software. We're basically moving from a system where people did a lot of work outside the system, a lot of co-ordination, a lot of chasing, and people managed the process. We are now shifting some of that work to the system. The system owns the process, the system delivers the outcome, people stay in charge of policies. They stay in charge of business objectives, they tell the system what they want to accomplish, and the system goes and gets it done. Opening the claw All of which brings us to today and Fusion Claw, pitched by Oracle as “a governed agentic execution runtime for Oracle Fusion Agentic Applications that combines AI reasoning with deterministic enterprise computation to enable highly complex work to be economically completed at scale” Initially there are 25 Claw-powered applications to add to the current tally of 75 Agentic Applications. The pitch for the latest enhancement is that Fusion Claw enables Agentic Applications to take on more complex, specialist-grade functions. As to how it works, as per the official blah blah: > Each Claw Outcome begins with a frontier model that reasons, plans, learns, and adapts. Fusion Claw then uses deterministic enterprise computation to execute precisely and at scale. This separation provides Fusion Applications customers with a potential economic advantage by applying AI reasoning only where intelligence is needed and using efficient deterministic computation for high-volume execution. > > > To govern autonomous enterprise execution, Fusion Claw applies an Enterprise Operating Envelope, which includes an organization’s objectives, standard operating procedures, policies, constraints, permissions, risk thresholds, decision rights, approval requirements, and escalation boundaries. The operating envelope is enforced across all outcomes through an Outcome Trust Harness that applies specific governing authority and controls to each outcome run, including identity, capabilities, data, and actions. > > > When the outcome is complete, an Outcome Receipt provides an auditable record of exactly what happened, including the authority applied, evidence used, decisions made, actions and transactions executed, and the result. In addition, Fusion Applications customers can decide what level of automation they want a process to have, ranging from quick assistance to governed full auto execution within explicitly delegated authority. Rachelson picks up the story: > Claw sits between the system of record and the agentic layer within Fusion. So, it powers Agentic Applications, but it also it communicates with the Agentic Applications as well as the system of record to get the work done. That's very very important. The interesting thing about Fusion Claw is it is a governed agentic execution runtime, and it performs one continuous outcome run. We designed it in a super-smart way, because we are using LLM (Large Language Model) planning and reasoning, the ability of LLM to analyze what is happening and come up with a plan. We then take the LLM offline, and the reconciliation, simulation, optimization, calculations, all of that happens in the other part of Fusion Claw. She notes: > That part is isolated from the LLM, she argues, and there are two reasons for this. One, Fusion Claw requires frontier models. Frontier models need capacity. Secondly, they're fairly expensive. So, there’s a timely tokenomics spin to all this as well, as she goes on: > What we're trying to do is use the LLM, which we absolutely have to, and work offline away from the LLM to do precise deterministic computation...In the agents that comprise Agentic Applications, each agent is granted certain authority, and a lot of the authority stems from the authority that's granted to a user today. So it adheres to all all of those requirements, does the computation. > > > What is really interesting is it is a self-learning and adaptive system. So once you come up with a plan, we keep to that plan, it is recorded, and we know what the plan is, but as conditions change, the plan just gets adjusted. But the plan does not need to get re-run, so you're not constantly using the expense of LLM to come up with a new plan. You are adapting the existing plan. You can also re-use the existing plan for other similar problems...There is quite a large proportion of work where the tokens do not get used. Business use cases Another notable claim being made for Fusion Claw is around its ability to support business needs. Rachelson says: > A business person gives the system a business objective, what needs to be solved, and the system does all its work, and it comes back with a business outcome. That business outcome can be a schedule for the hospital, a logistics plan for how you organize your trucks and which routes they take, or it can be a plan for how you fill the trucks, what kind of packages are gonna go in, sizes of packages, information about whether the packages need to be refrigerated or they can be stored at whatever temperature exists out in the universe. But the weather changed, Claw knows the weather changed, and so the day the weather changes, the routes get get modified. Other use cases might include sophisticated supply chains management, she suggests: > That planning function usually has a group called Research Operations. That group is staffed with highly trained, statisticians and and modeling researchers. Today, these researchers run old-style statistical models that are very complex and expensive to run. They use these statistical models to create simulations, scenario planning, taking myriad data sets and constraints into consideration and crunching through very large,, huge data sets. Fusion Claw really simplifies that work. She also turns to a favor